Bria releases Fibo 2.0 Dev, an 8B open-weights model that unifies image editing and generation

Bria AI has released Fibo 2.0 Dev, an 8B-parameter open-weights model that combines image editing and text-to-image generation in a single set of weights. According to Bria, it ranks #1 in image editing and #2 in text-to-image among open-weights models on the Artificial Analysis Image Arena, as of October 20, 2026.

What Fibo 2.0 Dev is

Fibo 2.0 is the next generation of Bria's FIBO family. Bria says it is the first FIBO release in which editing and generation share one model. Fibo 2.0 Dev is the full, non-distilled base model. It runs with classifier-free guidance and is positioned as the starting point for professional fine-tuning (LoRAs or full fine-tunes) and research. Bria lists Fibo 2.0 Edit Turbo, a distilled variant, for few-step, low-latency editing.

Like earlier FIBO models, Fibo 2.0 is trained on long, structured captions written in Visual GenAI Language (VGL), a JSON format. Bria says captions run up to 1,000+ words and describe lighting, composition, color, camera and depth of field as separate fields. Under this scheme, generation means rendering a structured blueprint, and editing means applying a change to that blueprint. The open-source bria-vgl package converts short prompts and plain-language edit instructions into structured prompts.

Workflow modes

Bria lists five modes:

• Edit: source image plus an instruction.
• Masked edit: source image, mask and instruction, with changes confined to the masked region.
• Multi-reference: up to 4 reference images plus an instruction, for composition, virtual try-on, product integration, and style or appearance transfer.
• Generate: a short text prompt that a VLM expands into a structured prompt.
• Inspire: a reference image plus a short prompt. The new structured prompt borrows the reference's subject, style or composition, and the model itself never sees the image.

Bria claims clear gains over previous FIBO releases in legible in-image text, material texture and human anatomy (hands, faces, poses). Text-to-image also offers optional style modes, each with its own prompts and negative prompt: Boring Reality, Kodak Film, Hard Flash, Hazy / Dreamy and Lens Distortion.

Licensing, data and access

Weights are released under Creative Commons Attribution-Non-Commercial (CC BY-NC 4.0). Commercial use requires a commercial agreement with Bria. Access to the weights on Hugging Face requires logging in and completing a form. Bria says the model is trained exclusively on licensed data.

Bria lists availability as open weights, API endpoints and ComfyUI nodes. On Fal.ai the endpoints are bria/fibo-2/generate and bria/fibo-2/edit. Hugging Face Spaces are available for text-to-image (Fibo-2.0-T2I) and editing (Fibo-2.0-editing, which runs Edit Turbo). Running the model locally requires diffusers>=0.42.0. Pricing, context window and numeric benchmark scores are not stated on the model page.

Announced next releases

Bria lists several upcoming releases built on Fibo 2.0: a Multi-Turn Local Editing LoRA for chained edits with the rest of the image kept pixel-stable, Background Removal, a Product Shot LoRA, Eraser by Mask, and Image Expansion (outpainting to any aspect ratio). No dates are given.

What this means

Fibo 2.0 Dev puts editing and text-to-image behind one deployment, which Bria says removes the need to run separate models for the two tasks. Because Dev is the non-distilled base with classifier-free guidance, it is aimed at teams that want to fine-tune for a specific domain, style or editing task rather than use it for low-latency production editing. That role is assigned to Edit Turbo.

The CC BY-NC 4.0 license limits the open weights to non-commercial use, so commercial deployment depends on an agreement with Bria. The #1 and #2 open-weights rankings are Bria's report of Artificial Analysis results dated October 20, 2026, and arena positions can change. The licensed-data claim is Bria's own statement about its training set.